Skip to content

opencode.md

Latest commit

 

History

History
96 lines (62 loc) · 4.92 KB

opencode.md

File metadata and controls

96 lines (62 loc) · 4.92 KB
Raw

OpenCode in Patchlab

Patchlab ships OpenCode as the built-in default tool. patchlab create . launches OpenCode in the sandbox without passing --tool.

Configuration sources

OpenCode settings come from two places:

  1. Your host OpenCode files (default on)

    • ~/.config/opencode/ — config, agents, plugins (opencode.json, tui.json, …)
    • ~/.local/share/opencode/auth.json — provider credentials from opencode auth login

    XDG paths are honored when XDG_CONFIG_HOME / XDG_DATA_HOME are set.

  2. ~/.patchlab/configuration.yaml (user-global overrides)

default_tool: opencode

tool_configuration:
  opencode:
    copy_host_configuration: true
    copy_host_auth: true
    proxy_local_models: true
    environment:
      SOME_FLAG: "1"

Patchlab copies host config into the sandbox before launch and rewrites loopback model URLs (see below). Host API keys (ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENCODE_*, …) are injected when set on the host.

When copy_host_auth is enabled (default), Patchlab also stages a copy of auth.json under the patchlab archive at ~/.patchlab/<id>/opencode-staging/ with restrictive permissions (0700 directories, 0600 files). That copy persists for the lifetime of the sandbox so patchlab resume can re-inject current credentials into the container. patchlab destroy <id> removes the archive (and staged auth) when branch cleanup succeeds.

Interactive launch and passthrough

Forward OpenCode CLI flags on any create/resume that launches the tool:

# TUI with a default model
patchlab create . --passthrough=--model --passthrough anthropic/claude-sonnet

# JSON output from a one-shot prompt
patchlab create . -p "Summarize README.md" \
  --passthrough=--format --passthrough json \
  --prompt-file ./README.md

Use --passthrough=--flag when the token starts with - (Commander otherwise treats it as another patchlab flag). --passthrough with --no-interactive (and no -p) is rejected — omit --no-interactive or use -p.

--prompt-file stages files into $HOME/context/ (same machinery as --context) and passes container paths to opencode run --file. It requires -p.

One-shot prompts

Run a single prompt without launching the TUI:

patchlab create . -p "Add unit tests for utils.ts"
patchlab create . -p          # read prompt from stdin (place -p last, or before other flags)
patchlab resume <id> -p "Fix the failing test"
patchlab resume <id> -p -     # explicit stdin sentinel

For prompts that start with -, use equals form: --prompt=-fix this.

Patchlab invokes opencode run --auto inside the sandbox (with --continue on resume). OpenCode --auto approves permissions that are not explicitly denied in your OpenCode config. The response is printed to stdout; patchlab status messages go to stderr. After the session is extracted, patchlab propagates OpenCode's exit code.

When the first pass produces no assistant text (for example, the model spends the turn on tools only), patchlab automatically runs a synthesis --continue prompt so you still get a written answer without manual follow-up.

--prompt cannot be combined with --no-interactive. Custom YAML providers opt in by declaring prompt_launch_command in the manifest (see configured-tool-provider.md). A YAML provider that replaces the built-in with name: opencode and overrides_builtin: true does not inherit the built-in's prompt argv or --passthrough unless it declares prompt_launch_command itself.

Local models (Ollama, LM Studio, …)

Models bound to 127.0.0.1 on the host are not reachable from inside the container as localhost. When opencode.proxy_local_models is true (default), Patchlab:

  1. Starts a detached host-side TCP proxy for each loopback port found in your OpenCode config
  2. Adds host.patchlab.internal to the container (--add-host)
  3. Rewrites http://localhost:PORT (and 127.0.0.1 / [::1]) in copied config to http://host.patchlab.internal:<listen-port>

Patchlab sets container HOME and USER to the provider image home (e.g. /home/patchlab) at sandbox create/resume so OpenCode reads copied config under ~/.config/opencode/ rather than the base image's default home (e.g. /home/node on node:22-slim).

The proxy keeps running after patchlab create exits (the sandbox container stays up). patchlab destroy <id> stops the proxy.

To disable the proxy (model must listen beyond loopback, e.g. OLLAMA_HOST=0.0.0.0:11434):

tool_configuration:
  opencode:
    proxy_local_models: false

Session resume

OpenCode project data under ~/.local/share/opencode/project/ inside the container is extracted on exit and restored on patchlab resume.

Overriding the built-in

Register a YAML provider with name: opencode and overrides_builtin: true to replace the built-in (see configuration-based-providers.md).